CSEPC: a deep learning framework for classifying small-sample multimodal medical image data in Alzheimer's disease

Area of Science:

• Medical Imaging Analysis
• Neurodegenerative Disease Research
• Machine Learning in Healthcare

Background:

• Alzheimer's disease (AD) poses a significant global health challenge, particularly for the elderly.
• Accurate staging of AD is critical for timely intervention and disease management.
• Limited sample sizes in medical imaging studies hinder the development of effective AD classification models.

Purpose of the Study:

• To introduce a novel diagnostic algorithm, cross-scale equilibrium pyramid coupling (CSEPC), designed for small-sample multimodal medical imaging data.
• To address the challenge of limited sample sizes in Alzheimer's disease (AD) classification.
• To improve the accuracy of AD diagnosis and staging.

Main Methods:

• Developed CSEPC, a novel algorithm integrating semantic features across different imaging modalities (sMRI, fMRI) and scales.
• Employed a cross-scale pyramid module for balanced multiscale feature extraction.
• Utilized contrastive learning-based cosine similarity coupling to capture intermodality associations and reduce parameters, optimizing for small sample sizes.

Main Results:

• The CSEPC model achieved high accuracy (85.67%) and AUC (0.98) in classifying mild cognitive impairment (MCI) to AD progression.
• Demonstrated superior performance in diagnosing different stages of Alzheimer's disease (AD).
• Outperformed existing models in AD classification tasks on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.

Conclusions:

• The CSEPC algorithm shows significant potential for small-sample multimodal medical imaging classification, especially for AD.
• This advancement can aid clinicians in managing and intervening in early-stage AD progression.
• The model's effectiveness in diagnosing and staging AD highlights its clinical utility.